Alibaba's desktop AI assistant QoderWork gets a 'consciousness' feature with memory and reflection

Alibaba's desktop AI assistant QoderWork just rolled out a 'consciousness' feature that combines memory, reflection, and skill evolution. It doesn't just save chats—it actively cleans up redundant info and turns repeated tasks into reusable skills.

The memory system works in layers: daily chats go into short-term memory, while valuable cross-session content moves to long-term memory. The reflection mechanism mimics how humans forget selectively—it takes a snapshot before each reflection, then uses hard rules to roll back if something goes wrong, so no key info is accidentally deleted. All memory is stored as open Markdown files on your local device, letting you view, edit, or delete them anytime.

When reflection is triggered, a multi-factor scoring model weighs user intent depth, chat length, and task complexity. Only when the score hits a threshold does the evolution process kick off. Evolution tasks fork from the main session into lightweight sub-sessions via a Fork Session mechanism, with over 99% cache hit rate and extra cost kept under 5% of the main session's total. If the system spots repeated tasks, it auto-generates skill suggestions—you can accept, ignore, or reject them, and rejected ones cool down without nagging again.